LEAST-TO-MOST PROMPTING ENABLES COMPLEX REASONING IN LARGE LANGUAGE MODELS

Abstract

Chain-of-thought prompting has demonstrated remarkable performance on various natural language reasoning tasks. However, it tends to perform poorly on tasks which requires solving problems harder than the exemplars shown in the prompts. To overcome this challenge of easy-to-hard generalization, we propose a novel prompting strategy, least-to-most prompting. The key idea in this strategy is to break down a complex problem into a series of simpler subproblems and then solve them in sequence. Solving each subproblem is facilitated by the answers to previously solved subproblems. Our experimental results on tasks related to symbolic manipulation, compositional generalization, and math reasoning reveal that least-to-most prompting is capable of generalizing to more difficult problems than those seen in the prompts. A notable finding is that when the GPT-3 code-davinci-002 model is used with least-to-most prompting, it can solve the compositional generalization benchmark SCAN in any split (including length split) with an accuracy of at least 99% using just 14 exemplars, compared to only 16% accuracy with chain-of-thought prompting. This is particularly noteworthy because neural-symbolic models in the literature that specialize in solving SCAN are trained on the entire training set containing over 15,000 examples. We have included prompts for all the tasks in the Appendix.

1. INTRODUCTION

Despite the great success of deep learning in the past decade, there still remain huge differences between human intelligence and machine learning: (1) Given a new task, humans usually can learn to accomplish it from only a few demonstration examples, while machine learning requires a large amount of labeled data for model training; (2) Humans can clearly explain the underlying rationale for their predictions or decisions, while machine learning is essentially a black box; (3) Humans can solve problems more difficult than any they have seen before, while for machine learning, examples in training and testing are typically at the same level of difficulty. The recently proposed chain-of-thought prompting approach (Wei et al., 2022; Chowdhery et al., 2022) has taken a significant step for narrowing the gap between human intelligence and machine intelligence. It combines the idea of natural language rationales (Ling et al., 2017; Cobbe et al., 2021) with few-shot prompting (Brown et al., 2020) . When further integrated with self-consistency decoding (Wang et al., 2022b) rather than using the typical greedy decoding, few-shot chain-of-thought prompting largely outperforms the state-of-the-art results in the literature on many challenging natural language processing tasks obtained from specially designed neural models trained with hundreds of times more annotated examples, while being fully interpretable. However, chain-of-thought prompting has a key limitation-it often performs poorly on tasks that require generalization of solving problems harder than the demonstration examples, such as compositional generalization (Lake & Baroni, 2018; Keysers et al., 2020) . To tackle such easy-to-hard generalization issues, we propose least-to-most prompting. It consists of two stages: first decomposing a complex problem into a list of easier subproblems, and then sequentially solving these

